FEANet: Frequency-Enhanced Attention Network with detail-preserving fusion for pulmonary disease classification

Introduction

Pulmonary diseases remain a leading cause of morbidity and mortality worldwide (1,2), including chronic obstructive pulmonary disease (COPD), lung cancer, and coronavirus disease 2019 (COVID-19)-related pulmonary infection. Medical imaging modalities such as computed tomography (CT), X-ray radiography, and magnetic resonance imaging (MRI) play a central role in clinical diagnosis (3). In practice, however, image interpretation still largely depends on radiologists, which is time-consuming and may yield inconsistent results due to inter-observer variability and the subtle visual differences among pulmonary pathologies. These challenges motivate the development of reliable computer-aided diagnosis systems for pulmonary disease classification.

Deep learning methods, particularly Convolutional Neural Networks (CNNs) (4-6) and Vision Transformers (ViTs) (7,8), have achieved strong performance in medical image classification. Nevertheless, most existing models mainly learn representations in the spatial domain and under-utilize frequency-domain cues related to texture, structural continuity, and edge sharpness, which are often indicative of thoracic abnormalities (9). Moreover, when training data are limited, spatial-domain models may become sensitive to modality-specific noise or acquisition artifacts (10). This sensitivity is influenced by the interaction between limited data volume and model complexity, rather than being unique to spatial-domain representations. Although ViTs capture long-range dependencies, they do not explicitly encode multi-scale frequency characteristics that are intrinsic to medical images (11). Therefore, improving robustness and generalization across modalities and acquisition conditions remains an open problem.

To address these limitations, recent studies have explored hybrid schemes that combine spatial- and frequency-domain information (12,13). In the frequency view, low-frequency (LF) components primarily describe global structure and semantics, whereas high-frequency (HF) components preserve edges and fine textures (14,15). Wavelet transform is particularly attractive due to its multi-resolution space-frequency localization (16,17). However, in many existing wavelet-based approaches, the wavelet transform is used mainly as a fixed input-level decomposition (18) or as a shallow auxiliary enhancement branch (19). In such designs, the decomposed frequency components are not sufficiently coupled with subsequent feature learning, attention allocation, or cross-frequency fusion. As a result, the potential of frequency cues to guide spatial representation learning may be underutilized, and HF details may still be attenuated during feature fusion.

In this work, we propose FEANet, a Frequency-Enhanced Attention Network with detail-preserving fusion for pulmonary disease classification. FEANet adopts a dual-branch design by applying discrete wavelet transform (DWT) to decompose an input image into LF and HF subbands. The LF branch employs a Wavelet-Aware Semantic Context Attention (WASCA) module to model global anatomical semantics via parallel frequency and spatial attention. The HF branch uses a Frequency-Guided Detail Attention (FGDA) module that learns content-adaptive frequency masks to highlight discriminative spectral components while refining local textures. Finally, a Frequency-Conditioned Fusion Attention (FCFA) module performs energy-aware cross-modal fusion, where HF discrete cosine transform (DCT) energy serves as an explicit attention bias to preserve frequency-rich details and alleviate texture blurring. The main contributions of this work are as follows:

• We introduce FEANet, a DWT-based dual-branch framework that jointly learns complementary LF semantics and HF details via dedicated encoders and cross-modal fusion.
• We design FGDA and WASCA to exploit frequency statistics for spatial feature learning, establishing an explicit correspondence between spectral importance and spatial attention.
• We propose FCFA, which injects HF details into LF features through energy-biased cross-attention and gated residual refinement, mitigating HF attenuation during fusion.
• Experiments on multiple pulmonary disease benchmarks (lung cancer, COPD, and COVID-19) demonstrate that FEANet achieves competitive accuracy and robust generalization across heterogeneous imaging settings.

We present this article in accordance with the STARD-AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0735/rc).

Methods

Datasets

We evaluated the proposed method on three datasets covering complementary imaging modalities, including two public benchmarks and one private clinical cohort.

• Lung cancer CT scan dataset: this public dataset contains 927 CT images from four categories—adenocarcinoma [314], large cell carcinoma [166], squamous cell carcinoma [245], and normal [202].
• COVID-19 radiography dataset: this public chest X-ray dataset includes four classes—COVID-19 [3,616], normal [10,192], pulmonary opacity [6,012], and viral pneumonia [1,345].
• COPD CT dataset: this private dataset was collected at Wuxi No. 2 People’s Hospital from patients who underwent pulmonary function testing between 2022 and 2025. Chest CT scans were performed using a Siemens SOMATOM go.Top scanner. Participants were scanned in the supine position during full inspiration, covering the lung apex to the diaphragm. Acquisition parameters were: tube voltage 70–140 kV with automatic tube current modulation (13–625 mA), 64-detector-row configuration with 1.5 mm slice collimation, gantry rotation time 1.0 s, pitch 1.0, and reconstruction matrix 512×512 pixels. Images were reconstructed using sinogram affirmed iterative reconstruction with attenuation values from −1,024 to 3,072 HU and stored in digital imaging and communications in medicine (DICOM) format. The dataset consisted of 162 patients who underwent both chest CT examination and pulmonary function testing within a clinically acceptable interval. COPD severity was determined according to the “Global Initiative for Chronic Obstructive Lung Disease” criteria based on post-bronchodilator spirometric measurements. Patients were categorized into Grade 1–4 according to the percentage of predicted forced expiratory volume in one second, while healthy subjects were assigned Grade 0. To construct the COPD dataset, 20–30 axial CT slices containing the largest visible lung regions were selected from each patient. These slices inherited the corresponding patient-level COPD severity label determined from pulmonary function testing. The selected slices were subsequently grouped according to patient identity before dataset partitioning, ensuring strict separation between training, validation, and test cohorts. The final dataset contained 3,624 Grade 0 images, 1,457 Grade 1 images, 2,099 Grade 2 images, 2,648 Grade 3 images, and 2,443 Grade 4 images. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Wuxi No. 2 People’s Hospital (No. WXEY-2026-86). All data were anonymized before analysis, and the requirement for informed consent was waived owing to the retrospective nature of the study.

Overall architecture

As shown in Figure 1, FEANet is a dual-branch framework that explicitly separates global semantics and fine-grained details using a single-level Haar DWT. Given an input image X, DWT produces a LF subband X_LF [low-low (LL)] and three HF subbands [low-high (LH), high-low (HL), high-high (HH)], concatenated as X_HF. The HF branch produces detail-preserving features F_HF using the FGDA module, while the LF branch extracts semantic features F_LF using the proposed WASCA module. The two streams are then integrated by the FCFA module via energy-aware cross-attention, yielding a fused representation X_O. Finally, X_O is fed into a classification head to produce the class logits. To evaluate the generalizability of FEANet across heterogeneous imaging modalities, a separate FEANet model was trained independently for each dataset. No joint training or parameter sharing was performed across CT and chest X-ray datasets. The reported results therefore reflect the ability of the proposed architecture to adapt to different imaging modalities under independent training settings. Implementation-level architectural details of FEANet are provided in Table S1.

Figure 1 Overview of the proposed FEANet framework. Conv, convolutional; FEANet, Frequency-Enhanced Attention Network; HF, high-frequency; HH, high-high subband; HL, high-low subband; LF, low-frequency; LH, low-high subband; LL, low-low subband.

HF branch: FGDA module

The HF subbands encapsulate fine-grained details including edge boundaries, texture patterns, and micro-structures. Different spatial regions within HF subbands possess distinct frequency characteristics—edge regions exhibit HF energy concentration, while homogeneous areas display low-to-mid frequency prevalence—yet existing methods fail to establish correspondence between frequency importance and spatial attention, resulting in inadequate emphasis on discriminative regions. To address this, we propose the FGDA module, which employs content-adaptive frequency mask learning to dynamically prioritize frequency-rich regions.

We first embed the HF input X_HF into a feature map:

H=GELU(GN(Conv1×1(XHF)))[1]

where GN(·) denotes group normalization, and GELU(·) denotes the Gaussian error linear unit activation.

As illustrated in Figure 2, in the proposed adaptive frequency mask generation module, we extract a global context descriptor from H:

S=GELU(Conv1×1(GAP(H)))[2]

Figure 2 Architecture of the proposed adaptive frequency mask generation module. Conv, convolutional; GELU, Gaussian error linear unit; GN, group normalization.

where GAP(·) denotes global average pooling over the spatial dimensions.

Given k heads, we adopt a multi-head mask generation module that takes the global context descriptor S as input and outputs k candidate frequency masks {Mi}i=1k together with their content-dependent aggregation weights {αi}i=1k (with ∑i=1kαi=1). The adaptive frequency mask is

Ms=∑i=1kαiMi[3]

We then modulate the block-wise DCT coefficients of H using the adaptive frequency mask M_S and frequency dropout D, and apply a block-wise inverse DCT to reconstruct a frequency-refined spatial representation:

Hr=IDCT(H⊙Ms⊙D)[4]

Hf=GELU(GN(Conv1×1(Hr)))[5]

where IDCT(·) denotes the block-wise inverse DCT, and ⊙ denotes the element-wise multiplication.

In parallel, we capture multi-scale textures from H using dilated depthwise convolutions:

U1=DWConv3×31(H)[6]

U2=DWConv3×33(H)[7]

U3=DWConv3×35(H)[8]

Hs=GN(Conv1×1(Concat[U1;U2;U3]))[9]

where Concat(·) denotes channel-wise concatenation, and DWConv3×3d(⋅) denotes a 3×3 depthwise convolution with dilation rate d.

We learn an adaptive gate from the concatenated outputs of both pathways and obtain the HF branch output via gated fusion:

g1=σ(GN(Conv1×1([Hf;Hs])))[10]

HHF=GELU(g1⊙Hf+(1−g1)⊙Hs)[11]

where σ(·) denotes the sigmoid activation function.

LF branch: WASCA module

The LF subband encodes global semantics, including coarse structural layout and LF intensity variations. However, conventional CNNs process spatial features uniformly without leveraging frequency-domain statistics to guide spatial learning, leading to insufficient exploitation of frequency information and missing critical directional characteristics inherent in medical images. To address these limitations, we propose the WASCA module, which employs a parallel dual-pathway design: a frequency attention module extracts ultra-low frequency statistics via block-wise DCT for channel-wise prioritization, while a spatial attention module uses Sobel-enhanced coordinate attention for directional localization.

First, the LF subband X_LF is embedded into a multi-scale feature representation via parallel dilated convolutions:

V1=Conv3×31(XLF)[12]

V2=Conv3×32(XLF)[13]

V3=Conv3×33(XLF)[14]

F=GELU(GN(Conv1×1(Concat([V1;V2;V3]))))[15]

where Conv3×3d(⋅) denotes a 3×3 convolution with dilation rate d.

Motivated by the observation that LF DCT coefficients capture coarse semantic trends, we propose a frequency attention module (Figure 3). Specifically, we partition F into non-overlapping blocks and apply a block-wise two-dimensional (2D) DCT to obtain the spectral representation:

Y=DCT(F)[16]

Figure 3 Architecture of the proposed frequency attention module. GELU, Gaussian error linear unit.

We then retain a predefined LF index set Ω and summarize the selected coefficients by spectral global average pooling (SGAP) to form a channel descriptor:

I=SGAP(Y)[17]

Specifically, SGAP first selects the DCT coefficients indexed by Ω from each block in Y, takes their magnitudes, and then averages them over all spatial blocks and the selected frequency indices, producing a compact per-channel spectral descriptor.

The descriptor I is mapped to channel-wise attention weights, which is subsequently used to re-weight the feature map:

Ff=F⊙σ(WbGELU(WaI))[18]

where W_a and W_b are learnable parameters.

To complement frequency statistics with directional localization, we introduce a spatial attention module that instantiates coordinate attention with Sobel-based directional priors. Given the feature map F, we use F(c,y,x) to denote its response at channel c and spatial location (y,x) (i.e., an equivalent element-wise notation introduced for convenience in the following derivations). We then compute axis-wise descriptors:

zH(c,x)=1H∑y=1HF(c,y,x)[19]

zW(c,y)=1W∑x=1WF(c,y,x)[20]

and obtain coordinate attention weights through a shared 1×1 transform followed by a split operation and two branch-specific projections:

t=GELU(BN(Conv1×1([zH;zW])))[21]

sH=σ(Conv1×1(SplitH(t)))[22]

sW=σ(Conv1×1(SplitW(t)))[23]

where BN(·) denotes batch normalization.

We further extract axis-wise Sobel cues by applying Sobel operators K_x and K_y:

mH=1HC∑y=1H∑c=1C|Ky∗F(c,y,x)|[24]

mW=1WC∑x=1W∑c=1C|Kx∗F(c,y,x)|[25]

where * denotes 2D convolution, |·| denotes element-wise absolute value.

We modulate F by combining coordinate attention with Sobel-derived directional cues:

Fs=F⊙σ(sH⊙(1+γmH)+sW⊙(1+γmW))[26]

where y controls the modulation strength.

Finally, we integrate the frequency-attended features F_f, the Sobel-enhanced spatial-attended features F_S, and the original features F by channel-wise concatenation, followed by a 1×1 convolution, normalization, and GELU activation:

FLF=F+GELU(GN(Conv1×1([Ff;Fs;F])))[27]

LF and HF feature fusion: FCFA module

Effective fusion of LF and HF features is critical. Conventional fusion strategies operate uniformly without preserving HF detail integrity, causing texture degradation and boundary blurring. We propose FCFA module, which preserves HF details through energy-aware cross-modal attention.

We first partition the spatial feature maps F_LF and H_HF into non-overlapping blocks and reshape each block into a token, resulting in two block-wise token sequences T_LF and T_HF. For each HF token, we reshape it back to its corresponding spatial block and perform a block-wise 2D DCT to obtain the DCT coefficients Z. Based on Z, we define the DCT energy-based bias as

En=1C∑cZ2[28]

bn=βlog(En+ϵ)[29]

where β is a scaling factor and ϵ is a small constant for numerical stability.

In the multi-head cross-attention with energy bias module (Figure 4), we use LF tokens as queries to retrieve detail-aware information from HF tokens. The query, key, and value projections are

Q=TLFWQ[30]

K=THFWK[31]

V=THFWV[32]

Figure 4 Architecture of the proposed multi-head cross-attention with energy bias module. HF, high-frequency; LF, low-frequency.

where W_Q, W_K, W_V are learnable matrices. The energy-biased cross-attention is computed as

Ta=(Softmax(QKTd+bn)V)WO[33]

where W_O is a learnable output projection, and d denotes the per-head feature dimension.

We map the attended tokens back to the spatial grid by assigning each token to its corresponding block location:

Rs=BU(MLP(Ta))[34]

where BU(·) denotes block unpatchify (token-to-block reshaping with block-wise broadcasting), and MLP(·) denotes a lightweight multi-layer perceptron applied to tokens.

Finally, we employ a residual gating module to regulate the injection of HF-enhanced features:

G=σ(GN(Conv1×1([FLF;Rs])))[35]

Xo=FLF+G⊙Rs[36]

Results

Image preprocessing

Prior to model training, all images underwent a unified preprocessing pipeline to ensure cross-dataset consistency. Specifically, all images were resized to 224×224 pixels to match the network input resolution and normalized to the range of [0,1].

For the public lung cancer CT dataset, the released data were provided as preprocessed image files rather than raw DICOM volumes. Therefore, no Hounsfield Unit (HU) conversion or DICOM-specific processing was required. Images were directly resized and normalized before being fed into the model.

For the private COPD CT dataset, original chest CT scans were acquired in DICOM format and converted into axial slices. Intensity values were first clipped using a lung window setting of [−1,000, 400] and then linearly normalized to [0, 1] using min − max scaling. To ensure sufficient pulmonary parenchyma coverage, 20–30 representative axial slices were selected per patient based on maximal lung region visibility.

For the COVID-19 chest X-ray dataset, all radiographs were uniformly resized to 224×224 pixels and normalized to [0, 1], consistent with the preprocessing strategy applied to CT data.

Experimental settings

All experiments were conducted on Ubuntu 22.04 using the PyTorch 2.1.2 framework and an NVIDIA RTX 3090 GPU with 24 GB memory. The datasets were divided into training, validation, and test sets with a ratio of 7:2:1. The training and validation subsets were used for model optimization and hyperparameter tuning, while the independent test sets were reserved for final performance evaluation. The model was trained using the Adam optimizer with a learning rate of 10⁻⁴ and a batch size of 32, together with an early stopping strategy to mitigate overfitting. Given the moderate dataset size, mild data augmentation was applied to CT slices during training, including random rotations (±10°), in-plane translations (±5%), horizontal flipping, and slight brightness and contrast adjustments. Model performance was evaluated using accuracy, AUC, precision, recall, and F1-score. All experimental results reported in this study were obtained on the independent test sets.

All baseline models were trained and evaluated under identical experimental settings to ensure a fair comparison. Specifically, the same training, validation, and test splits, optimization strategy, learning rate schedule, batch size, early stopping criterion, and data augmentation procedures were used for all methods. For models with publicly available pretrained parameters (DenseNet121, VGG16, ViT-B, and PVT-V2), ImageNet-pretrained weights were used for initialization, followed by end-to-end fine-tuning on each target dataset. Models without publicly available pretrained weights were trained using their officially released implementations and recommended configurations. Hyperparameters were tuned using the validation set under the same protocol for all competing methods.

Evaluation metrics

Model performance was evaluated using five standard classification metrics: accuracy, precision, recall, F1, and the area under the receiver operating characteristic curve (AUC). All metrics were computed based on the confusion matrix, including true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN).

Accuracy measures the proportion of correctly classified samples:

Accuracy=TP+TNTP+TN+FP+FN[37]

Recall evaluates the proportion of positive samples correctly identified:

Recall=TPTP+FN[38]

Precision reflects the reliability of positive predictions:

Precision=TPTP+FP[39]

F1 is the harmonic mean of precision and recall:

F1=2×Precision×RecallPrecision+Recall[40]

AUC represents the area under the ROC curve and provides a threshold-independent measure of classification performance.

Comparison with deep learning methods

To evaluate the effectiveness of the proposed FEANet, we conducted comparative experiments against a diverse set of representative deep learning models, including conventional CNNs (DenseNet121 and VGG16), Transformer-based architectures (ViT-B and PVT-V2), recent lung disease classification methods [MCNN-AM (20) and Lung-EffNet (21)], and frequency-enhanced models based on wavelet transforms [WAAtt (22) and WACNN (23)]. All models were trained and evaluated using the same data splits and optimization settings to ensure a fair comparison.

Results on lung cancer CT scan dataset

Table 1 and Figure 5 show that FEANet achieves the best overall performance on the lung cancer CT dataset. FEANet attains an accuracy of 0.9870 and an AUC of 0.9995, outperforming the strongest competing method by +1.30 percentage points in accuracy and +0.14 points in AUC. These results indicate that explicitly disentangling LF semantics and HF details, followed by frequency-conditioned fusion, is beneficial for lung cancer subtype recognition.

Figure 5 Confusion matrix on the lung cancer CT dataset. CT, computed tomography; DenseNet121, 121-layer Dense Convolutional Network; FEANet, Frequency-Enhanced Attention Network; M, metastasis; MCNN-AM, Multi-Scale Convolutional Neural Network with Attention Module; N, node; PVT-V2, Pyramid Vision Transformer version 2; T, tumor; VGG16, Visual Geometry Group 16-layer Convolutional Neural Network; ViT, Vision Transformer; WAAtt, Wavelet Attention; WACNN, Wavelet Attention Convolutional Neural Network.

Results on COVID-19 radiography dataset

As reported in Table 2, FEANet yields the highest accuracy (0.9604) and F1 (0.9604) despite the pronounced class imbalance. Compared with the strongest baseline, FEANet improves accuracy and F1 by +0.52 percentage points and achieves the highest AUC (0.9951). The confusion matrix in Figure 6 further shows stable recognition across classes, suggesting that FEANet captures discriminative radiographic patterns while remaining robust to inter-class similarity.

Figure 6 Confusion matrix on the COVID-19 radiography dataset. COVID-19, coronavirus disease 2019; DenseNet121, 121-layer Dense Convolutional Network; FEANet, Frequency-Enhanced Attention Network; MCNN-AM, Multi-Scale Convolutional Neural Network with Attention Module; PVT-V2, Pyramid Vision Transformer version 2; VGG16, Visual Geometry Group 16-layer Convolutional Neural Network; ViT, Vision Transformer; WAAtt, Wavelet Attention; WACNN, Wavelet Attention Convolutional Neural Network.

Results on COPD CT dataset

Results on the clinical COPD CT dataset are summarized in Table 3. FEANet achieves the best performance across all metrics, reaching 0.9984 accuracy and 0.9998 AUC. Although multiple methods perform strongly on this dataset, FEANet still provides a consistent improvement over the best competing approach, indicating strong generalization to real-world clinical CT data. The confusion matrix in Figure 7 shows highly consistent predictions across disease severity grades.

Figure 7 Confusion matrix on the COPD CT dataset. COPD, chronic obstructive pulmonary disease; CT, computed tomography; DenseNet121, 121-layer Dense Convolutional Network; FEANet, Frequency-Enhanced Attention Network; MCNN-AM, Multi-Scale Convolutional Neural Network with Attention Module; PVT-V2, Pyramid Vision Transformer version 2; VGG16, Visual Geometry Group 16-layer Convolutional Neural Network; ViT, Vision Transformer; WAAtt, Wavelet Attention; WACNN, Wavelet Attention Convolutional Neural Network.

Ablation analysis

To quantify the contribution of each proposed component, we performed a systematic ablation study on the COVID-19 radiography dataset. The results are summarized in Table 4. The baseline model uses the same backbone and DWT decomposition as FEANet but removes all frequency-guided attention and fusion designs. Adding WASCA yields a clear improvement, increasing accuracy from 0.9472 to 0.9546 and AUC from 0.9896 to 0.9918, indicating that LF semantic context modeling enhances global representation learning under class imbalance. Introducing FGDA alone also improves performance (95.21% accuracy; 0.9909 AUC), highlighting the importance of selectively emphasizing discriminative HF textures and edge cues. In contrast, using FCFA alone provides only marginal gains (95.05% accuracy; 0.9903 AUC), suggesting that frequency-conditioned fusion becomes most effective when the fused features are already enriched by frequency-aware representations. Combining modules leads to larger and more consistent improvements. In particular, jointly enabling WASCA + FGDA substantially boosts performance (95.89% accuracy; 0.9936 AUC), demonstrating the complementary roles of LF semantic guidance and HF detail modeling. Incorporating FCFA on top of these encoders further improves both accuracy and AUC, and the full model FEANet achieves the best results, reaching 96.04% accuracy, F1 and 0.9951 AUC.

Visualization of frequency-guided attention maps

To qualitatively examine how frequency cues guide spatial representation learning for COPD severity grading, we visualize intermediate outputs of FEANet on representative CT slices from Grades 0–4. As shown in Figure 8, each row corresponds to a COPD grade and includes: the original CT image, the LF spatial attention map produced by WASCA, the adaptive HF mask learned by FGDA in the frequency domain, and the final fused attention map generated by FCFA.

Figure 8 Visualization of frequency-guided attention maps. COPD, chronic obstructive pulmonary disease; CT, computed tomography; FCFA, Frequency-Conditioned Fusion Attention; FGDA, Frequency-Guided Detail Attention; WASCA, Wavelet-Aware Semantic Context Attention.

Overall, the WASCA attention maps consistently highlight the lung fields across all grades, indicating that LF information effectively captures global anatomy and coarse intensity distributions. However, these responses are relatively smooth, with limited sensitivity to localized structural variations that may correlate with disease severity. In contrast, the FGDA masks exhibit complementary, block-wise activation patterns in the frequency domain. The activated frequency components vary across grades, suggesting that FGDA performs content-adaptive frequency selection rather than indiscriminate enhancement. This behavior is consistent with severity-dependent changes in fine-grained lung textures. Finally, the FCFA fused attention maps present more concentrated and structured spatial responses than WASCA alone. By conditioning spatial attention on the frequency importance learned by FGDA, FCFA suppresses irrelevant background regions while emphasizing lung areas with pronounced textural irregularities. Notably, as COPD severity increases, the fused attention becomes progressively more heterogeneous, which aligns with the expected progression of emphysematous changes. These visualizations support the key design premise of FEANet: frequency-domain cues can explicitly guide spatial attention allocation to yield spatially meaningful, detail-sensitive representations.

Robustness analysis

We further evaluated the robustness of FEANet under common image degradations on the COPD CT test set, including Gaussian blur, Gaussian noise, salt-and-pepper noise, and resolution downsampling. Blur severity was controlled by the Gaussian kernel standard deviation (q), noise severity by the pixel-value standard deviation (p), and downsampling by the scale factor (s). These perturbations emulate some clinical challenges such as motion blur, sensor noise, impulsive artifacts, and reduced acquisition resolution. Quantitative results are summarized in Table 5. FEANet exhibits strong robustness under mild-to-moderate perturbations, while performance degrades as corruption severity increases. On clean images, FEANet achieves an accuracy of 0.9984 with an AUC of 0.9998. Under Gaussian blur, accuracy decreases from 0.9862 at q=2 to 0.8605 at q=4, and further to 0.6188 at q=6 (AUC: 0.9997 → 0.9791 → 0.8692), indicating that severe blur substantially disrupts discriminative structural cues. A similar trend is observed for Gaussian noise: accuracy drops from 0.9813 at p=10 to 0.7810 at p=20, and to 0.4647 at p=30 (AUC: 0.9995 → 0.9683 → 0.8362). For salt-and-pepper noise, FEANet remains highly stable at low corruption (0.9943 accuracy at 0.5%), but performance declines at higher noise levels (0.8751 at 1% and 0.6650 at 1.5%). Notably, FEANet is particularly robust to resolution degradation: even with aggressive downsampling (s=0.3), accuracy remains 0.9895 and AUC remains 0.9996, suggesting that the proposed frequency-enhanced representations preserve discriminative information under reduced spatial resolution.

Discussion

This study presents FEANet, a Frequency-Enhanced Attention Network for pulmonary disease classification across heterogeneous imaging modalities. By explicitly integrating frequency-domain cues with spatial representation learning, FEANet targets a key limitation of conventional CNN- and Transformer-based classifiers that predominantly rely on spatial-domain features. Experiments on three pulmonary imaging datasets demonstrate that FEANet achieves consistently strong performance, with high accuracy and AUC across both public benchmarks and a real-world clinical cohort.

Our results suggest that frequency information provides complementary and modality-resilient cues for pulmonary disease recognition. In our comparisons, Transformer-based models (e.g., ViT-B and PVT-V2) underperform on CT-based tasks, which may be partially attributed to limited sensitivity to subtle texture variations and boundary patterns in data-scarce settings. CNN baselines exhibit stronger robustness but can still be influenced by local appearance bias. In contrast, FEANet decomposes input images into LF structural components and HF detail components via wavelet analysis, and learns to fuse them through frequency-guided attention. This design promotes more stable representations across CT and X-ray modalities, which is further supported by qualitative analyses showing that frequency-aware attention progressively focuses on heterogeneous lung regions with increasing disease severity.

The high classification performance observed on the COPD dataset may be attributed to several factors. First, all COPD severity labels were derived from clinically established pulmonary function measurements, providing reliable patient-level annotations. Second, representative CT slices with substantial lung-region coverage were selected, reducing the influence of non-informative anatomical structures. Finally, advanced COPD stages often exhibit pronounced structural abnormalities, including emphysematous destruction and heterogeneous lung attenuation patterns, which may facilitate discrimination among severity grades.

Compared with prior wavelet-based approaches that treat wavelet decomposition as a preprocessing step or apply shallow/late fusion, FEANet incorporates frequency information as a first-class signal throughout the attention learning pipeline. Specifically, the explicit coupling between frequency importance and spatial attention, implemented through energy-aware conditioning and gated residual fusion, helps preserve fine-grained textures and mitigates HF attenuation that commonly occurs in naive fusion schemes. The robustness experiments further corroborate this behavior: FEANet remains highly stable under moderate perturbations and exhibits strong tolerance to resolution degradation, indicating that frequency-enhanced representations can retain discriminative information even when spatial details are partially compromised.

Recent advances in medical foundation models and domain generalization have demonstrated strong potential for improving robustness across heterogeneous imaging environments. Large-scale pretrained models, including vision-language foundation models and medical image foundation models, leverage extensive datasets to learn transferable representations and have achieved promising performance across diverse downstream tasks. Similarly, domain generalization approaches aim to improve model robustness under distribution shifts caused by differences in imaging devices, acquisition protocols, and patient populations. Compared with these approaches, FEANet focuses on an orthogonal perspective by explicitly modeling frequency-domain information and integrating it with spatial representations through frequency-guided attention mechanisms. The proposed framework does not rely on large-scale pretraining and can be trained directly on task-specific datasets. Future work will investigate the integration of frequency-aware representation learning with foundation-model architectures and domain generalization strategies to further enhance cross-domain robustness and clinical applicability.

Despite these promising results, several limitations should be acknowledged. First, the present study used a fixed train-validation-test partition rather than repeated k-fold cross-validation. Although the independent test sets remained completely unseen during model development and identical data partitions were used for all competing methods, a single holdout evaluation cannot fully characterize the statistical variability associated with dataset partitioning. Future studies should therefore employ repeated cross-validation and external validation using larger, multicenter cohorts to provide more comprehensive estimates of performance uncertainty and generalizability. Second, although FEANet generalizes well across CT and X-ray, its applicability to other modalities (e.g., MRI or ultrasound) remains to be validated. Third, frequency-domain operations introduce additional computational cost compared with standard CNN baselines; future work will explore more efficient implementations and lightweight variants for time-critical deployment. Finally, while attention visualizations provide intuitive interpretability, their clinical relevance has not been formally validated. Future studies involving expert radiologists and lesion-level annotations are needed to assess whether the highlighted regions align with clinically meaningful patterns and to further strengthen the interpretability claims.

Conclusions

In this work, we proposed FEANet, a Frequency-Enhanced Attention Network that jointly exploits frequency-domain cues and spatial representations for pulmonary disease classification. Extensive experiments on CT and chest X-ray datasets demonstrate that FEANet achieves competitive accuracy and robust generalization across heterogeneous imaging settings. In future work, we will investigate extensions to additional modalities, develop more efficient variants, and further validate the clinical relevance of the learned attention patterns with expert assessment.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the STARD-AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0735/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0735/dss

Funding: This work was supported by Medical Key Strategic Project of Wuxi Health Commission.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0735/coif). This research was supported by the Medical Key Strategic Project of Wuxi Health Commission and Wuxi Municipal Bureau on Science and Technology (No. Y20242206). The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Wuxi No. 2 People’s Hospital (No. WXEY-2026-86). All data were anonymized before analysis, and the requirement for informed consent was waived owing to the retrospective nature of the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Wang Z, Lin J, Liang L, Huang F, Yao X, Peng K, Gao Y, Zheng J. Global, regional, and national burden of chronic obstructive pulmonary disease and its attributable risk factors from 1990 to 2021: an analysis for the Global Burden of Disease Study 2021. Respir Res 2025;26:2. [Crossref] [PubMed]
2. Park SL, Lichtensztajn D, Yang J, Wu J, Shariff-Marco S, Stram DO, Inamdar P, Fruin S, Larson T, Tseng C, Setiawan VW, Gomez SL, Samet J, Le Marchand L, Wilkens LR, Ritz B, Wu AH, Cheng I. Ambient Air Pollution and Chronic Obstructive Pulmonary Disease: The Multiethnic Cohort Study. Ann Am Thorac Soc 2025;22:698-706. [Crossref] [PubMed]
3. Kwong RY, Yucel EK. Computed tomography scan and magnetic resonance imaging. Circulation 2003;108:e104-6. [Crossref] [PubMed]
4. Ahmad IS, Dai J, Xie Y, Liang X. Deep learning models for CT image classification: a comprehensive literature review. Quant Imaging Med Surg 2025;15:962-1011. [Crossref] [PubMed]
5. Yadav SS, Jadhav SM. Deep convolutional neural network based medical image classification for disease diagnosis. J Big Data 2019;6:113.
6. Han Q, Qian X, Xu H, Wu K, Meng L, Qiu Z, Weng T, Zhou B, Gao X DM-CNN. Dynamic Multi-scale Convolutional Neural Network with uncertainty quantification for medical image classification. Comput Biol Med 2024;168:107758. [Crossref] [PubMed]
7. Okolo GI, Katsigiannis S, Ramzan N. IEViT: An enhanced vision transformer architecture for chest X-ray image classification. Comput Methods Programs Biomed 2022;226:107141. [Crossref] [PubMed]
8. Manzari ON, Ahmadabadi H, Kashiani H, Shokouhi SB, Ayatollahi A. MedViT: A robust vision transformer for generalized medical image classification. Comput Biol Med 2023;157:106791. [Crossref] [PubMed]
9. Qiong L, Chaofan L, Jinnan T, Liping C, Jianxiang S. Medical image segmentation based on frequency domain decomposition SVD linear attention. Sci Rep 2025;15:2833. [Crossref] [PubMed]
10. Tang S, Ran H, Yang S, Wang Z, Li W, Li H, Meng Z. A frequency selection network for medical image segmentation. Heliyon 2024;10:e35698. [Crossref] [PubMed]
11. Raghu M, Unterthiner T, Kornblith S, Zhang C, Dosovitskiy A. Do vision transformers see like convolutional neural networks? Adv Neural Inf Process Syst 2021;34:12116-28.
12. Cai M, Zhang H, Huang H, Geng Q, Li Y, Huang G. Frequency domain image translation: more photo-realistic, better identity-preserving. 2021 IEEE/CVF International Conference on Computer Vision (ICCV); Montreal, QC, Canada. 2021:13910-20.
13. Lin S, Zhang Z, Huang Z, Lu Y, Lan C, Chu P, You Q, Wang J, Liu Z, Parulkar A. Deep frequency filtering for domain generalization. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); Vancouver, BC, Canada. 2023:11797-807.
14. Wang Y, Wang X, Zhang S, Huang Z, Liu S, Feng L. WFA-SRNet: a wavelet-guided and feature-aware network for remote sensing image super-resolution. IEEE Trans Geosci Remote Sens 2025;1-14.
15. Yang W, Chen B, Shen Y, Yu L. WaveCNNs-AT: wavelet-based deep CNNs of adaptive threshold for signal recognition. Appl Intell 2023;53:28819-31.
16. Khatami A, Khosravi A, Nguyen T, Lim CP, Nahavandi S. Medical image analysis using wavelet transform and deep belief networks. Expert Syst Appl 2017;86:190-8.
17. Gao F, Fu M, Cao J, Dong J, Du Q. Adaptive frequency enhancement network for remote sensing image semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing 2025;1-15.
18. Chen G, Sun F, Jing W, Zou W, Di D, Song Y, Fan L. Learning frequency-domain fusion for multimodal remote sensing semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing 2025;1-16.
19. Tang Y, Xie X, Yu Y. Hyperspectral classification of two-branch joint networks based on Gaussian pyramid multiscale and wavelet transform. IEEE Access 2022;10:56876-87.
20. Thakare V, Aote SS, Raipurkar A. A novel framework for lung disease classification using multiscale convolutional neural networks with an integrated dynamic attention mechanism. Int J Imaging Syst Technol 2025;35:e70155.
21. Raza R, Zulfiqar F, Khan MO, Arif M, Alvi A, Iftikhar MA, Alam T. Lung-EffNet: lung cancer classification using EfficientNet from CT-scan images. Eng Appl Artif Intell 2023;126:106902.
22. Zhao X, Huang P, Shu X. Wavelet-attention CNN for image classification. Multimed Syst 2022;28:915-24.
23. Rakhimov M, Karimberdiyev J, Javliev S. Artificial intelligence in medicine: enhancing pneumonia detection using wavelet transform. Intelligent Human Computer Interaction: 15th International Conference, IHCI 2023, Daegu, South Korea, November 8–10, 2023:165-75.